Large scale graph processing systems: survey and an experimental - - PowerPoint PPT Presentation

Large scale graph processing systems: survey and an experimental evaluation

Cluster Computing 2015

Omar Batarfi, Radwa El Shawi, Ayman G. Fayoumi , Reza Nouri, Seyed-Mehdi-Reza Beheshti, Ahmed Barnawi, Sherif Sakr

Graph scale

• Scale: millions to billions of nodes and edges
• Facebook social network graph:
• 1 billion+ users (nodes)
• 140 billion+ friendship relationships (edges)
• The size of a single node/edge can be very different due to the various attributes size or nodes/edges
• An Estimation of large-scale graph size: 1 GB to 500 TB
• ~0.5 GB/million nodes, ~0.20 GB/million edges
• 1 million nodes/edges: ~0.5 GB nodes & ~0.20 GB edges
• 500 billion nodes/edges: ~250 TB nodes & ~100 TB edges

GB/million nodes GB/million edges 0.42 0.20 0.84 0.13 11.39 2.60 1.83 0.07 0.58 0.23 0.48 0.19 0.44 0.18

platforms

• General-purpose platforms (such as MapReduce) are bad
• No direct support for iterative graph algorithms
• To detect a fix point (termination condition), an extra task might be required on each iteration
• Specialized platforms
• Pregel family
• Graphlab family
• Others

Algorithms

• Characteristics of graph algorithms
• Iterative
• Need to traverse the graph
• Typical graph algorithms
• PageRank: rank nodes based on their incoming/outgoing edges
• Shortest Path: find the path between 2 nodes where the sum of weights is minimal
• Pattern Matching: find certain structures (e.g. path, star)
• Triangle Count: counts the number of triangles
• Connected Component: find the subgraphs in which any two vertices are connected

Pregel family

• Pregel:
• Google’s pioneer work in this area
• Published in 2010
• Distributed & computations are totally in memory
• Iteration -> superstep
• Address scalability issue
• Bulk Synchronous Parallel (BSP): synchronization barrier on each superstep
• Message passing interface (MPI)
• Vertex-centric approach
• Locality: Each vertex & its neighbors are in the same node
• A vertex can: execute a function/send messages to others/change states (active/inactive)
• Termination: no active vertices & no messages being transmitted
• Pregel family
• Apache Giraph: Java implementation of pregel
• GPS: another Java implementation
• Pregelix: set-oriented, iterative dataflow

GraphLab family

• GraphLab
• Shared memory
• GAS (Gather, Apply, Scatter) processing model
• Gather: a vertex collects info of its neighbors
• Apply: performs computation
• Scatter: update adjacent vertices and edges
• Comparison
• GAS : pull-based; a vertex request info of all neighbors
• MPI: push-base; a vertex receives info from neighbors
• Two modes:
• Synchronous model (BSP): communication barriers
• Asynchronous model: using distributed locking; no communication barriers orsuperstep
• GraphLab family
• PowerGraph: avoid the imbalanced workload caused by high degree vertices in power-law graphs
• Trinity: memory-based; distributed
• Signal/Collect: vertex-centric; two operations for a vertex (signal/collect)
• Graphchi

GraphLab family cont’d

• Graphchi:
• Out-of-core: using secondary storage in a single machine
• Parallel Sliding Window (PSW):
• Goal: decreases non-sequential accesses on disk
• It partitions the graph into shards
• In each shard, edges are sorted by the source IDs
• Selective scheduling:
• Converge faster on some parts of the graph
• “some parts” -> the change on values is significant
• Pros
• It avoids the challenge of finding efficient graph cuts
• Now with zone-based devices, partitioning is needed again
• It avoids cluster management, fault tolerance etc.
• Out-of-Core + SMR

Other systems

• TurboGraph
• Out-of-core
• Processing billion-scale graphs using modern hardware -> parallelism
• Multicore CPU: multiple job at the same time
• FlashSSD: multiple I/O requests in parallel using multiple flash memory packages
• A parallel model called pin-and-slide: column view of the matrix-vector multiplication
• Two types of thread pools
• Execution thread pool
• Asynchronous I/O callback thread pool
• Steps
• Restrict computation to a set of vertice -> identify the corresponding pages
• -> pin those pages in the buffer pool
• -> processing completes for a page -> swtich unpinned -> can be evicted now
• -> Parallel asynchronous I/O request to the FlashSSD for pages which are not in the buffer pool
• The system can slide the processing window one page at a time
• Multiple channel SSD
• Extreme-large-scale graph that does fit into memory
• CMR -> SMR/Zone named SSD

Other systems

• GRACE
• Out-of-core
• Batch-style graph programming frameworks
• Providing a high level representation for graph data
• Separating application logic from execution policies.
• Combine synchronous programming with asynchronous execution

Experiments

• Perforamance metrics
• Reading Time, Processing Time, Writing Time, Total Execution Time, CPU Utilization, RAM Usage, Network Traffic
• Deployed on Amazon AWS cloud services

The execution times metrics for the PrageRank algorithm for all systems using the different datasets